test(acl): cover who may create a data set, and by which route

[JosteinGj]

The suite proved a per-data-set write grant cannot create a data set, and nothing else. That leaves three gaps.

• No allow side. A rule that only ever denies is indistinguishable from one that denies everyone: the suite would stay green through a change that made data sets uncreatable, with the failures landing in unrelated fixtures where they read as fixture breakage.
• Only the near-miss principal. MT_WRITEONLY holds a write grant, just a scoped one. The plain cases - read-only and no-grant - are what would catch a check keyed on "holds any write grant" rather than on the blanket one.
• Only one of the two routes. /resources/create gates DATASET/POLICY nodes with its own check, so proving /datasets/create is guarded says nothing about it. That route is the one the typed node surface makes natural - resources.create(vec![Dataset::new(..)]).

Three Rust tests and one Python one. The Python side stays deliberately thin, per that file's own selection principle: it covers the /resources route because that is a different binding path (the typed node union), not because the backend rule needs re-proving.

Two things the writing found, both now pinned in place rather than in a comment somewhere:

• The lowercase-label case is deliberate. A type-label is canonicalised before it is matched, so a gate comparing the raw string would let dataset through.
• A policy create needs its type. Bean validation runs before the ACL gate, so a policy without one is a 400 naming nodes[0].type and never reaches the permission check - a test built on an invalid body would pass against a backend with no gate at all.

Rust: 17 ACL tests pass (was 14). Python: 10 (was 9).